can't update specular in principled bsdf using pytorch optimization.

[anewusername77]

Hi there again.
i read pytorch intergate tutorial already. and I tried as follows:

import mitsuba as mi
import drjit as dr
mi.set_variant("llvm_ad_rgb")

envmap = mi.Bitmap()
light = mi.load_dict({
    'type': 'envmap',
    'bitmap': envmap
})
scene =  mi.load_dict({
            "type": "scene",
            "integrator": {
                "type": "path",
                'max_depth': 8,
                'hide_emitters':True},
            "light": light,
            "face": {
                'filename': 'face.ply',
                'type': 'ply',
                'bsdf': {
                        'type': 'principled',
                        'base_color': {
                            'filename': 'diff2.png',
                            'type': 'bitmap',
                            },
                        'specular':0.0,
                        'roughness': 0.0,
                    },
            }
            })
origin = T.rotate([0, 0, 1], 90).rotate([0, 1, 0], 180) @ mi.ScalarPoint3f([0, 0, 500])   
cam = mi.load_dict({
            'type': 'perspective',
            'to_world': T.look_at(
                    origin=origin,
                    target=[0, 0, 0],
                    up=[0, 0, 1]
                ),
            
          'fov': 10.,
          'sampler': {
             'type': 'independent',
            'sample_count': 16
          },
         'film': {
                    'type': 'hdrfilm',
                    'width':  512,
                     'height': 512,
                    'rfilter': {
                        'type': 'tent',
                    },
                    'pixel_format': 'rgba',
                 },
})

@dr.wrap_ad(source='torch', target='drjit')
def renderimg(scene, vertices:torch.Tensor, faces:torch.Tensor, normals:torch.Tensor, 
         uvmap:torch.Tensor, albedo_tex:torch.Tensor, specular:torch.Tensor, 
         roughness:torch.Tensor, envmap:torch.Tensor, fov, sensor, samples):
       
        params_scene = mi.traverse(scene)
        
        params_scene['light.data'] = mi.TensorXf(envmap)
        params_scene.update()
        # mesh
        vertex_texcoords = mi.Point2f(mi.Float(uvmap[:,0].numpy()), 
                            mi.Float(uvmap[:,1].numpy()))
        params_scene['bfm.vertex_positions'] = dr.ravel(mi.TensorXf(vertices))
        params_scene['bfm.vertex_normals'] = dr.ravel(mi.TensorXf(normals))
        params_scene['bfm.faces'] = dr.ravel(mi.TensorXf(faces))
        params_scene['bfm.vertex_texcoords'] = dr.ravel(vertex_texcoords)
       
        ## bsdf
        params_scene['bfm.bsdf.base_color.data'] = albedo_tex
        params_scene['bfm.bsdf.roughness.value'] = dr.ravel(roughness)
        params_scene['bfm.bsdf.specular'] = dr.ravel(specular)
        
        params_scene.update()
        
        params_sensor = mi.traverse(sensor)
        params_sensor['x_fov'] = dr.ravel(fov)
        return mi.render(scene, sensor=sensor, spp=samples, seed=1)
...
optimizer = torch.optim.Adam([
            {'params': vShCoeffs, 'lr': 0.005},
            {'params': specular, 'lr': 0.001}
               ....
        ])

loss = l1loss(img_gt, rendered_img)

although the rendered_img.requires_grad==True, specular.requires_grad==True, the params I want to optimize like specular, fov, albedo_tex, vertex_positions didn't update at all.

[njroussel]

I think this another case where you're not passing params to mi.render, or am I missing something?

[anewusername77]

yes, thank you so much, already solved in issue. ( don't know how to delete this discussion

[njroussel]

Nevermind, I thought this was an issue and not a discussion 😅

Good news that you found it!